Skip to content
theodor-semerdzhievPublic

About

Team Name: Reverse Algo

Resources

Stars

1 star

Watchers

0 watching

Forks

Latest commit

 

History

54 Commits

Folders and files

Repository files navigation

FIAM 2025 — Reverse Algo Submission

This is the project that was submitted for the FIAM 2025 Hackathaon for our team Reverse Algo. This project got us 4th place at FIAM. FIAM 2025 Official Site

Overview

Reverse Algo developed an ensemble trading architecture tailored for small-cap equities, integrating complementary signals that capture distinct dimensions of market behavior.

The system combines four independent sub-models:

  • Text-Based Sentiment & Risk-Tone Models — extract linguistic cues around risk, optimism, and earnings-surprise context.
  • Regime-Clustering Module — identifies shifts in volatility, liquidity, and momentum environments.
  • Factor-Momentum Pipeline — captures style rotation across value, quality, reversal, and momentum factors.
  • Historical Similarity Engine — uses vector similarity search to match current market states to historical precedents.

Outputs from these modules feed into a four-learner meta-model, with each learner transforming the feature set differently.
The meta-model dynamically reweights its internal components based on validation performance, enabling real-time adaptation to evolving market regimes.


Trading Strategy

The portfolio operated as a 70% long / 30% short book with 150/50 leverage, allocating exposures based on:

  • signal strength
  • conviction thresholds
  • dynamic risk scoring
  • portfolio constraints
  • number of qualifying securities

Trades are picked at the beginning of each month, based on the data on the previous months. This ensured capital-efficient scaling while avoiding overexposure or overfitting to transient signals.


Performance Summary

During the full test period, the strategy delivered:

  • Annualized Return: 34%
  • Sharpe Ratio: 0.93
  • CAPM Alpha: 16%
  • Beta: 1.26
  • Information Ratio: 0.56
  • Directional Accuracy (meta-model): 51.1%
  • Max Drawdown: –0.42
  • Monthly Turnover: 0.78
  • Small-Cap Exposure: 98%
  • Annualized Fee Drag: 18%

Performance was especially strong post-2020, when persistent momentum regimes favored the ensemble’s dynamic weighting and adaptive allocation processes.


Model Architecture

System Diagram

Text and quantitative data flow into four specialized sub-models: Risk Narrative, Management Discussion (Surprise), Regime Clustering, and Similarity Search. Their outputs combine with engineered features like factor momentum, profitability, value–quality interactions, and price-path metrics.

All signals are then fed into The Committee, an ensemble of Ridge, XGBoost, CATBoost, and QLattice, which produces the final predicted return used for trading.

Architecture Diagram

Equity Curve

Equity Curve from 2017 to 2025 compared to the S&P500 and MSCI World Index. Equity Curve

Authors

  • Samuel Beadoin
  • Theodor Semerdzhiev
  • Carter Cameron
  • Luyun Chen
  • Nicholas Charland Armstrong

About

Team Name: Reverse Algo

Resources

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages